Why you can't trust cloud API endpoints: 40 days of coding-agent telemetry showing the exact day providers silently overwrote deepseek-v4.1-flash with -1017 ("faster execution, shorter answers" = 3x more swearing and 6.7x more model switches)
retained text
Around mid-October `deepseek-v4.1-flash` suddenly started feeling much dumber in my coding agent, so I parsed my local session logs to see when it broke. Before the update it had a funny quirk where it would randomly reply in Chinese while being 100% convinced it was speaking English. In the `-1017` update (released Oct 17) they fixed the Chinese leak and made it fire tool calls more aggressively - and somehow lobotomized the model in the process. Providers silently overwrote the `deepseek-v4.1-flash` endpoint with the new weights by Oct 19. My swearing rate tripled, I ended up switching models mid-session 6.7x more often, and because it's a cloud API, there's no way to roll back to the old weights.
document fetched 2026-10-30 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xiqw9i/why_you_cant_trust_cloud_api_endpoints_40_days_of